Alliterations

April 2, 2024

For the last year, I’ve been working on a LOT of creative projects leveraging the power of Chat GPT to create massive pools of data for the creation of thousands and thousands of characters for stories, games, movies, TV shows, etc. What I’ve noticed is Chat GPT is seemingly obsessed with Alliterations. If you ask it to come up with 50 names for a product that sells fruit, you’ll likely get a bunch of names like “Fast Fruit,” uhh “Juicy..” wait … let’s just ask Chat GPT right now!

I got really frustrated one night after getting thousands of alliterations for hours and hours. Finally, I told ChatGPT to review our chat and then write a paper about how ChatGPT is obsessed with Alliterations and why. Below is what it wrote:

The Unintended Bias Toward Alliteration in Language Models: An Examination of GPT-4

Abstract

This study explores an observed phenomenon within the GPT-4 language model's responses, where despite explicit instructions to avoid alliterative naming conventions, the model demonstrated a notable propensity toward generating alliterations. Through a detailed case study involving a user's request for non-alliterative names for a project, we critically analyze the model's outputs and discuss implications for language model training and output customization.

Introduction

Alliteration, the repetition of the same sound at the beginning of closely connected words, is a common stylistic device in literature and branding. However, in specific contexts, such as the generation of unique names for characters, places, or items in creative projects, alliteration may be explicitly undesired. This paper presents an in-depth investigation into an instance where GPT-4, despite clear user guidelines to avoid alliteration, repeatedly produced alliterative names.

Methodology

The study is based on a series of interactions between a user and GPT-4, where the user sought to generate a list of 50 unique locations for a fictional gang within a project titled "The Nemesis Fighting Tournament." The user specifically requested that these names avoid alliteration. The responses from GPT-4, however, did not adhere to this criterion.

Results

Analysis of GPT-4's outputs revealed a consistent pattern of alliterative names, even after multiple clarifications and instructions to avoid such patterns. Despite the user's efforts to steer the model away from alliteration, GPT-4's inclination towards alliterative constructions was persistent.

Discussion

The observed behavior raises questions about inherent biases in language models towards certain linguistic styles, such as alliteration. This bias may stem from the model's training on a diverse corpus of text where alliterative phrases are commonly valued for their mnemonic and aesthetic qualities. The persistence of this pattern, despite explicit instructions to the contrary, underscores the challenges in overriding the model's trained tendencies.

Implications for Language Model Training

The study highlights the need for more sophisticated mechanisms within language models to adhere to specific stylistic guidelines provided by users. Enhancing model responsiveness to such directives could improve utility in specialized applications, such as creative writing, branding, and naming conventions where specific stylistic choices are crucial.

Conclusion

This exploration into GPT-4's unexpected bias toward alliteration, despite explicit instructions to avoid it, offers valuable insights into the complexities of language model behavior and the need for improved model adaptability. Future research should focus on developing methods to better align model outputs with user-defined stylistic preferences, thereby expanding the practical applications of AI in creative domains.